文献速递腹部CT

基于CT深度学习模型早期预测ESBL阳性肝脓肿

CT-based Deep Learning Model for Automatic Segmentation and Early Predicting of Pyogenic Liver Abscess Caused by Extended-Spectrum β-lactamase-Producing Enterobacteriaceae: A Multicenter Retrospective Study (CLASS2401).
PubMed 文献 · Acad Radiol · 2026-07-10

这项多中心回顾性研究纳入442例化脓性肝脓肿患者,开发并验证了基于nnUNetv2的自动分割及临床-影像联合模型。临床-影像模型在训练、内部测试和外部测试集中的AUC分别为0.945、0.889和0.844。该模型可在病原学确认前早期预测产超广谱β-内酰胺酶肠杆菌所致肝脓肿。

要点

  • 临床-影像联合模型在外部测试集AUC为0.844,可用于早期识别ESBL-PLA高危患者
  • 最佳截断值0.663时,外部测试集阳性预测值70.7%、阴性预测值95.7%
  • 模型预测的高危组ESBL-PLA发生率为74.58%,显著高于低危组的6.00%

核心结论

  • 基于深度学习的影像组学可在病原学确认前实现ESBL-PLA的早期影像预测
  • 临床-影像联合模型对ESBL-PLA具有较高预测效能,外部测试集AUC为0.844
  • 模型可为多重耐药病原体感染的影像学风险分层提供支持

研究设计与样本

  • 回顾性多中心研究,纳入中国6个中心共442例化脓性肝脓肿患者
  • 数据集划分为训练集211例、内部测试集111例、外部测试集120例
  • 基于nnUNetv2开发自动分割模型,由专家标注病灶
  • 构建临床模型、影像组学模型、深度学习影像模型以及临床-影像组学模型、临床-影像联合模型

关键影像表现/征象

  • 自动分割模型在训练、内部测试和外部测试集中均达到较高的Dice相似系数
  • 临床-影像联合模型在训练、内部测试和外部测试集的AUC分别为0.945、0.889和0.844

诊断标准与关键数值

  • 最佳截断值为0.663时,临床-影像联合模型在训练、内部测试和外部测试集的阳性预测值分别为95.2%、86.6%和70.7%
  • 最佳截断值为0.663时,临床-影像联合模型在训练、内部测试和外部测试集的阴性预测值分别为95.0%、96.3%和95.7%
  • 模型预测的高危组ESBL-PLA发生率为74.58%,低危组为6.00%,差异具有统计学意义(P < 0.001)

临床应用与阅片要点

  • 临床-影像联合模型可在微生物学确认前早期预测ESBL-PLA,有助于指导初始经验性抗菌治疗选择
  • CT影像特征结合临床信息可用于化脓性肝脓肿患者的耐药病原体风险分层

局限性

  • 研究为回顾性设计

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

英文原文(摘要)· 点击展开

RATIONALE AND OBJECTIVE: Pyogenic liver abscess (PLA) caused by extended-spectrum β-lactamase-producing Enterobacteriaceae (ESBL-PLA) presents major antimicrobial treatment challenges and delayed pathogen identification. This study developed and validated deep learning-based models integrating clinical and CT imaging features for early prediction of ESBL-PLA prior to microbiological confirmation.

MATERIALS AND METHODS: This retrospective multicenter study included 442 patients from 6 centers in China, divided into training, internal test, and external test sets (n = 211, 111, and 120, respectively). An automated PLA segmentation model based on nnUNetv2 was developed using expert-annotated lesions. Clinical, radiomics, and deep learning imaging features were used to construct predictive basic models. Two combined models, the clinical-radiomics model and clinical-imaging model (CIM) were constructed. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and other metrics.

RESULTS: The segmentation model achieved high Dice similarity coefficient values across datasets. The CIM showed strong predictive performance for ESBL-PLA, with AUCs of 0.945, 0.889, and 0.844 in the training, internal test, and external test sets, respectively. At the optimal cutoff (0.663), the CIM achieved positive predictive values of 95.2%, 86.6%, and 70.7%, and negative predictive values of 95.0%, 96.3%, and 95.7% in the training, internal test, and external test sets, respectively. And the high-risk group had a significantly higher incidence of ESBL-PLA (74.58%) than the low-risk group (6.00%, P < 0.001).

CONCLUSIONS: Deep learning-based radiomics enables early imaging-based prediction of ESBL-PLA and may support imaging-based risk stratification for infections caused by multidrug-resistant pathogens.

来源:Acad Radiol,查看原文。版权归原作者所有。

引用

  1. Gong, Wang, Guo, et al. CT-based Deep Learning Model for Automatic Segmentation and Early Predicting of Pyogenic Liver Abscess Caused by Extended-Spectrum β-lactamase-Producing Enterobacteriaceae: A Multicenter Retrospective Study (CLASS2401). Acad Radiol. 2026. PMID: 42431794. doi:10.1016/j.acra.2026.06.028

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。